AttributeError: 'collections.OrderedDict' object has no attribute 'eval'

Question

I have a model file which looks like this

OrderedDict([('inp.conv1.conv.weight', 
          (0 ,0 ,0 ,.,.) = 
           -1.5073e-01  6.4760e-02  1.9156e-01
            1.2175e-01  3.5886e-02  1.3992e-01
           -1.5903e-01  8.2055e-02  1.7820e-01
          
          (0 ,0 ,1 ,.,.) = 
            1.0604e-01 -1.3653e-01  1.4803e-01
            6.0276e-02 -1.4674e-02  2.3059e-06
           -6.2192e-02 -5.1061e-03 -7.4145e-03
          
          (0 ,0 ,2 ,.,.) = 
           -5.5632e-02  3.5326e-02  6.5108e-02
            1.1411e-01 -4.4160e-02  8.2610e-02
            8.9979e-02 -3.5454e-02  4.2549e-02
          
          (1 ,0 ,0 ,.,.) = 
            4.8523e-02 -4.3961e-02  5.3614e-02
           -1.2644e-01  1.2777e-01  8.9547e-02
            3.8392e-02  2.7016e-02 -1.4552e-01
          
          (1 ,0 ,1 ,.,.) = 
            9.5537e-02  2.8748e-02  3.9772e-02
           -6.2410e-02  1.1264e-01  7.8663e-02
           -2.6374e-02  1.4401e-01 -1.7109e-01
          
          (1 ,0 ,2 ,.,.) = 
            5.1791e-02 -1.6388e-01 -1.7605e-01
            3.5028e-02  7.7164e-02 -1.4499e-01
           -2.9189e-02  2.7064e-03 -2.3228e-02
          
          (2 ,0 ,0 ,.,.) = 
           -7.4446e-03 -9.7202e-02 -1.4704e-01
           -1.0019e-02  8.1780e-02 -5.3530e-02
           -1.8412e-01  1.5988e-01 -1.3450e-01
          
          (2 ,0 ,1 ,.,.) = 
           -1.1075e-01 -5.2478e-02  6.0658e-02
            1.6739e-01 -2.9360e-02  1.2621e-01
            2.0686e-02  1.1468e-01  1.2282e-01

I want to do inference on this model, but when i do model.eval() i get,

AttributeError: 'collections.OrderedDict' object has no attribute 'eval

score 48 · Accepted Answer · answered Apr 20 '18 at 13:14

It is not a model file, instead, this is a state file. In a model file, the complete model is stored, whereas in a state file only the parameters are stored.
So, your OrderedDict are just values for your model. You will need to create the model and then need to load these values into your model. So, the process will be something in form of

import torch
import torch.nn as nn

class TempModel(nn.Module):
    def __init__(self):
        self.conv1 = nn.Conv2d(3, 5, (3, 3))
    def forward(self, inp):
        return self.conv1(inp)

model = TempModel()
model.load_state_dict(torch.load(file_path))
model.eval()

You'll need to define your model properly. The one given in the example above is just a dummy. If you construct your model yourself, you might need to update the keys of the saved dict file as mentioned here. The best course of action is to define your model in exactly the same way from when the state_dict was saved and then directly executing model.load_state_dict will work.

Is there a way to save the model so that it wont need redefinition when used? — havakok, Oct 14 '18 at 10:58
You can take a look at this thread - https://stackoverflow.com/questions/42703500/best-way-to-save-a-trained-model-in-pytorch — layog, Oct 14 '18 at 11:06

score 4 · Answer 2 · answered Mar 05 '20 at 10:00

4

Simply use:

torch.save(model_conv,'cnn.pt')
the_model = torch.load('cnn.pt')

answered Mar 05 '20 at 10:00

Sijin John

492
7
15

3

The author has got the OrderedDict structure. This approach is correct in case when one has the whole model stored, not the parameters only, and wants to load it. – pdaawr Dec 06 '20 at 14:00

AttributeError: 'collections.OrderedDict' object has no attribute 'eval'

2 Answers2

Linked